|Publication number||US5901306 A|
|Application number||US 08/881,721|
|Publication date||May 4, 1999|
|Filing date||Jun 23, 1997|
|Priority date||Jun 23, 1997|
|Publication number||08881721, 881721, US 5901306 A, US 5901306A, US-A-5901306, US5901306 A, US5901306A|
|Inventors||Vladimir Y. Volkonsky|
|Original Assignee||Sun Microsystems, Inc.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (9), Non-Patent Citations (2), Referenced by (1), Classifications (13), Legal Events (6)|
|External Links: USPTO, USPTO Assignment, Espacenet|
1. Field of the Invention
This invention relates to a method and apparatus for optimizing overflow checking and reduction of data signals represented as signed 8-bit integers.
2. Description of Related Art
The explosion of graphics, audio, and video ("multimedia") related applications in computer systems has fueled efforts in improving processor efficiency with regard to processing multimedia signals. Multimedia signals include audio and pixel ("picture elements") signals, among other things, and which may be sufficiently represented using binary data of no more than eight bits of resolution. Binary data having greater widths may also be used but are often limited to intermediate results for advanced data manipulation since such data formats lead to an increase load on instruction execution, resulting in slower rates of data manipulation by a processor.
A computer system running a video application may represent color pixels through four signed 8-bit signals, with each signed 8-bit signal representing the three primary color values of red, green, and blue and an intensity value. This translates to a large amount of data signals required to represent an image for display on a computer screen even when restricting pixel data widths to eight bits. For example, to display a digital NTSC video signal in real-time on a computer monitor requires a pixel rate of 10.4 million pixels per second. With three data signals to manipulate per pixel, this translates to about 30 million pieces of data to manipulate per second. A processor clock rate of 200 million MHz would only have 20 clock cycles available for processing each pixel which is less than seven clock cycles for each primary color value.
Manipulating pixel data that are represented using signed 8-bit integers usually requires that the resulting pixel data remain within the maximum negative and positive boundaries of a signed 8-bit integer. A signed 8-bit integer in two's compliment format has a maximum range boundary of 127 and a minimum range boundary of -128. For example, when scaling or rotating images, it is necessary to combine the incoming signal being processed with other internally generated signal data in order to obtain the resulting pixel data. This ensures that if an overflow state does occur, the resulting pixel data is reduced to a value supported by the data format in which the pixel is represented. If either of the range boundaries is exceeded by the resulting pixel data, the resulting pixel data is reduced to within the maximum or minimum range boundaries of 127 and -128, respectively.
In the past, checking resulting pixel data for an overflow condition included using conditional branches. For example, in one such method branch operations in the programming language "C" are used in the following manner.
The resulting pixel data, which is represented as the variable "dst," is compared with the upper range boundary of 127. If the resulting pixel data exceeds the upper range boundary, then the upper range boundary value is transferred into the resulting pixel data. Otherwise, a conditional branch occurs which bypasses the execution of the second operation. The third operation compares the resulting pixel data with the lower range boundary of -128. If the resulting pixel data exceeds the lower range boundary, then the lower range boundary value is transferred into the resulting pixel data. Otherwise, another conditional branch occurs which bypasses the execution of the third operation, i.e., the resulting pixel data falls within the range boundaries.
The use of conditional branches in a superscalar pipelined processor decreases processor execution throughput because the branches interrupt the pipeline processing of instructions. Also, conditional branches usually require processors to perform a memory fetch from intermediate or main memory in the event of a cache miss. Since intermediate or main memory is typically much slower than an instruction register which is used to process the instructions, the time to process the conditional branches takes much longer to complete than instructions that do not require fetches from intermediate or main memory. Thus, not only does the processor incur an increase in fetch latency but it also takes an efficiency hit due to the fact that the pipelining of instructions has been interrupted by the branches.
Accordingly, it would be desirable to provide a method and apparatus that ensures resulting pixel data or similar data using signed 8-bit integers remain within the range boundaries of a signed 8-bit integer without the use of conditional branches in an instruction. This advantage is achieved by performing three shift operations, two logic multiplications and one logic addition to obtain a result that is within the range boundaries of a signed 8-bit integer, improving the instruction throughput of a processor.
The present invention is directed to checking and reducing an intermediate result signal arising from a manipulation of data signals without using conditional branches, thereby improving instruction processing in a superscalar pipelined processor. In the preferred embodiment of the present invention, the data signals are represented as signed 8-bit binary values in a two's compliment format. This requires that the intermediate result signal be stored in a register that is greater than 8-bits wide to allow for the proper checking of an overflow condition.
It is presently contemplated that the present invention include using a processor operating under program control with the program having the following operations. The program determines whether the intermediate result signal is in a positive overflow state or a negative overflow state. The program sets a first mask signal to have 8 lower bits in an OFF position when the intermediate result signal is inside the range of a signed 8-bit integer. Otherwise, the program sets the first mask signal to have 8 lower bits in an ON position. Also, the program sets a second mask signal to have 8 lower bits in the OFF position when the intermediate result signal is not in the positive or negative overflow state. Otherwise, it sets the second mask signal equal to an upper threshold signal when the data signal is in an overflow state or setting the second mask signal to a lower threshold signal when the data signal is in a negative overflow state. Finally, the program bitwise ANDs the intermediate result with the first mask signal to obtain a translated data signal, and bitwise ORs the translated data signal with the second mask signal.
FIG. 1 is a schematic block diagram illustrating a computer system.
FIG. 2 is a schematic block diagram illustrating a processor used in accordance with a preferred embodiment of the present invention.
FIG. 3 is a process flow diagram showing the method of operation in accordance with a preferred embodiment of the present invention.
In the following description, a preferred embodiment of the invention is described with regard to preferred process steps and data structures. Those skilled in the art would recognize after perusal of this application that embodiments of the invention can be implemented using one or more general purpose processors operating under program control, or special purpose processors adapted to particular process steps and data structures, and that implementation of the process steps and data structures described herein would not require undue experimentation or further invention.
The present invention is directed to checking and if applicable, reducing an intermediate result signal arising from a manipulation of data signals without using conditional branches, thereby improving instruction processing in a superscalar pipelined processor.
The data signals are represented as 8-bit signed integers which requires the intermediate result signal to be stored in a register that is greater than 8-bits wide. This allows for the proper checking of an overflow condition. In the preferred embodiment, the focus of whether an overflow condition exists focuses on the state of the 8th and 9th bits of the intermediate result data. Consequently, intermediate registers need to have an 8th and 9th bit position but may be conventionally sized as 16 bit or 32 bit fixed-data values to fit within the processing scheme of a processor.
Data manipulations resulting in intermediate results that require checking for an overflow condition arises from many different types of signal processing applications. Images may be super-imposed, requiring pixel signals from the image sources to be processed so that when displayed, one image is semi-transparent over the other image. If the images are comprised of pixels represented as 8-bit signed integers, this results in combining pixels that occupy the same pixel space to a single intermediate result signal. To ensure that the intermediate result signal is within the 8-bit signed integer data scheme, its overflow condition must be checked. And if in the overflow condition, reduced to within the range boundaries of a signed 8-bit integer. The range boundaries of a signed 8-bit integer in a two's compliment format include a maximum positive range boundary of 127 and a maximum negative range boundary of -128.
FIG. 1 is a schematic block diagram illustrating a computer system 8 in which the presently preferred invention would have application. Computer system includes a superscalar pipelined processor 10, an I/O adapter 12 coupled to a data store 14, a display adapter 16, a ROM 18, and RAM devices 20. A data and addressing bus 22 is also shown coupled to each of the items included with the computer system.
FIG. 2 is a schematic block diagram illustrating an example superscalar pipelined processor 24 used in accordance with a preferred embodiment of the present invention. A Prefetch and Dispatch Unit (PDU) 26, an Integer Execution Unit IEU) 28, a Floating-Point Unit (FPU) 30, a Memory Management Unit (MMU) 32, a Load and Store Unit (LSU) 34, an External Cache Unit (ECU) 36, a Graphics Unit (GRU) 38, an Instruction Cache 40, and a Data Cache 42 are shown in FIG. 2. Superscalar pipelined processors are known in the art of computer architecture. Consequently, those of ordinary skill in the art will readily recognize that processor 24 is capable performing arithmetic and logical operations that include bitwise shifting, bitwise AND operations, bitwise OR operations, and addition operations when operating under program control.
The PDU 26 ensures that all execution units remain busy by fetching instructions before they are needed in the pipeline. Instructions can be prefetched from all levels of the memory hierarchy, including instruction cache 40, external cache 36, and main memory. The PDU 26 provides a 12-entry prefetch buffer which minimizes pipeline stalls. In addition, the PDU 26 has a 9-stage instruction pipeline to minimize latency and dynamic branch prediction to allow for greater prediction accuracy.
The pipeline is a double-instruction-issue pipeline with nine stages: fetch, decode, grouping, execution, cache access, load miss, integer pipe wait, trap resolution, and writeback. These stages imply that the latency (time from start to end of execution) of most instructions is nine clock cycles. However, at any given time, as many as nine instructions can execute simultaneously, producing an overall rate of execution of one clock per instruction in many cases. However, some instructions may require more than one cycle to execute due to the nature of the instruction such as a branch instruction or to a cache miss, or other resource contention.
The first stage of the pipeline is a fetch from instruction cache 40. In the second stage, instructions are decoded and placed in the instruction buffer. The third stage, grouping, groups and dispatches up to four instructions. Next, integer instructions are executed and virtual addresses calculated during the execution stage. In the fifth stage data cache 42 is accessed. Cache hits and misses are determined, and branches are resolved. If a cache miss was detected, the loaded miss enters the load buffer. At this point, the integer pipe waits for the floating-point/graphics pipe to fill and traps are resolved. In the final stage, writeback, all results are written to the register files and instructions are committed.
IEU 28 includes two ALU (arithmetic logical units) for arithmetic, logical, and shift operations, an eight window register file, result bypassing, and a Completion Unit which allows a nine-stage pipeline with minimal bypasses.
FPU 30 is a pipelined floating-point processor that consists of five separate functional units to support floating-point and multimedia operations. The separation of execution units allows the issuance and execution of two floating-point instructions per cycle. Source and data results are stored in a 32-entry register file in either 8, 16 or 32 bit lengths. Most floating-point instructions have a throughput of one cycle, a latency of three cycles, and are fully pipelined. The FPU is able to operate on both single precision (32-bit), and double-precision (64-bit) numbers, normalized or denormalized, in hardware, and quad-precision (128-bit) operands in software.
FPU 30 is tightly coupled to the integer pipeline and is capable of seamlessly executing a floating-point memory event and a floating-point operation. IEU 28 and FPU 30 have a dedicated control interface which includes the dispatch of operations fetched by the PDU 26 to the FPU 30. Once in the queue,the PDU 26 is responsible for distribution of instructions to the FPU 30. IEU 28 controls the data cache portion of the operation, while the FPU 30 decides how to manipulate the data. The IEU 28 and FPU 30 cooperatively detect floating-point data dependencies. The interface also includes IEU 28 and FPU 30 handshaking for floating-point exceptions. The FPU 30 performs all floating-point operations and implements a 3-entry floating-point instruction queue to reduce the impact of bottlenecks at the IEU 28 and improve overall performance.
MMU 32 handles all memory operations as well as arbitration between data stores and memory.
GRU 38 relies on integer registers of varying bit lengths for addressing image data, and floating point registers for manipulating that data. This division of labor enables processor to make full use of all available internal registers so as to maximize graphical throughput.
Method of Operation
FIG. 3 is a process flow diagram showing the method of operation in accordance with a preferred embodiment of the present invention.
Referring now to FIG. 3, at step 44 a processor determines whether an intermediate result signal is in a positive overflow state or negative overflow state Specifically, the processor is directed under program control to bitwise shift the intermediate data signal four bit positions to the right which is then bitwise ANDed ("masked") with a signal having a value 18 (hex) to obtain a shift signal.
As known in the art, a bitwise AND operation passes through all bits that are ON in bit positions that correspond to bit positions in first mask signal having an ON bit. All other bit positions, whether ON or OFF, that correspond to bit positions in the first mask signal having an OFF bit are set to the OFF position.
As known in the art, shifting a binary value four positions to the right is equivalent to performing a divide operation where the dividend is equal to the value of 16. Also, the value 18 (hex) may be appropriately filled with additional binary signals of the value zero, depending on the size of the intermediate register used. This provides a level of scalability as to the size of the intermediate register.
At step 46, the processor shifts a first mask signal having a binary value of FF0000FF (hex) a number of bit positions to the right equal to the shift signal from step 44 to obtain a first mask template signal.
At step 48, the processor shifts a second mask signal having a binary value of 00807F00 (hex) a number of bit positions to the right equal to shift signal from step 44 to obtain a second mask template signal.
At step 50, the processor masks the intermediate result signal with first mask signal to obtain a translated data signal. Specifically, this mask operation includes bitwise ANDing the intermediate result signal with the first mask signal.
At step 52, the processor performs a bitwise OR operation on translated data signal using second mask signal. As known in the art, a bitwise OR operation turns all bits that are OFF to ON in bit positions that correspond to bit positions in second mask signal having an ON bit. All other bit positions, whether ON or OFF, that correspond to bit positions in the second mask signal having an OFF bit are passed through without change.
The above method of operation essentially is based on the analysis of the 9th and 8th bits of the intermediate result signal. The bits are interpreted in the following way. A binary status of "00" or "11" indicates that intermediate result signal is inside the range of a signed 8-bit integer in two's compliment format, which is a range between 127 and -128. A binary status of "01" indicates a positive overflow state with the intermediate result signal being greater than an upper boundary of 127. A binary status of "10" indicates a negative overflow state with the intermediate result signal being greater than a lower boundary of -128.
First and second masks signals are initialized to have initial values FF0000FF (hex) and 00807F00 (hex), respectively, which are bitwise shifted in response to steps 46 and 48. In essence, the first mask signal is set to have 8 lower bits in an OFF position when intermediate result signal is in the positive or negative overflow state. Otherwise first mask signal is set to have 8 lower bits in an ON position.
The second mask signal is set to have 8 lower bits in the OFF position when intermediate result signal is not in the positive or negative overflow state, i.e., when the intermediate result signal is inside the range of a signed 8-bit integer. Otherwise, the second mask signal is set equal to upper threshold signal when the intermediate result signal is in an overflow state or set to a lower threshold signal when the intermediate result signal is in a negative overflow state. For a signed 8-bit integer in two's compliment format the upper and lower threshold signals are equal to 127 and -128, respectively.
Alternative embodiments of the present invention may include embedding the steps of the method of the present invention into a Field Programmable Gate Architecture (FPGA) as is well-known in the art, or using an integrated circuit design program such as VHDL to describe the method, thus hard coding the method in an application-specific integrated circuit (ASIC). The skill necessary to perform such embedding and hard coding is well-known to those of ordinary skill in the art.
While preferred embodiments are disclosed herein, many variations are possible which remain within the concept and scope of the invention, and these variations would become clear to one of ordinary skill in the art after perusal of the specification, drawings and claims herein.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4675725 *||Aug 21, 1984||Jun 23, 1987||Crosfield Electronics Limited||Image superimposing system|
|US4945507 *||Jun 12, 1989||Jul 31, 1990||Nec Corporation||Overflow correction circuit|
|US5222159 *||Jun 19, 1991||Jun 22, 1993||Canon Kabushiki Kaisha||Image processing method and apparatus for extracting a portion of image data|
|US5402368 *||Dec 9, 1993||Mar 28, 1995||Fujitsu Limited||Computing unit and digital signal processor using the same|
|US5402506 *||Jun 23, 1994||Mar 28, 1995||Pixel Semiconductor, Inc.||Apparatus for quantizing pixel information to an output video display space|
|US5408670 *||Dec 18, 1992||Apr 18, 1995||Xerox Corporation||Performing arithmetic in parallel on composite operands with packed multi-bit components|
|US5508951 *||Nov 10, 1994||Apr 16, 1996||Matsushita Electric Industrial Co., Ltd.||Arithmetic apparatus with overflow correction means|
|US5539685 *||Aug 10, 1993||Jul 23, 1996||Kabushiki Kaisha Toshiba||Multiplier device with overflow detection function|
|US5734744 *||Jun 7, 1995||Mar 31, 1998||Pixar||Method and apparatus for compression and decompression of color data|
|1||"The UltraSPARC Processor--Technology White Paper; The UltraSPARC Architecture," pp. 1-10, Copyright 1994-1997 Sun Microsystems, Inc., Palo Alto, CA.|
|2||*||The UltraSPARC Processor Technology White Paper; The UltraSPARC Architecture, pp. 1 10, Copyright 1994 1997 Sun Microsystems, Inc., Palo Alto, CA.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US20100047331 *||Apr 11, 2008||Feb 25, 2010||Haynes Barton F||Method of inducing neutralizing antibodies to human immunodeficiency virus|
|U.S. Classification||712/224, 712/E09.069, 712/E09.02|
|International Classification||G06F9/30, G06F7/48, G06F9/38|
|Cooperative Classification||G06F9/3885, G06F9/30021, G06F7/48, G06F7/4991|
|European Classification||G06F9/30A1C, G06F9/38S, G06F7/48|
|May 15, 1998||AS||Assignment|
Owner name: SUN MICROSYSTEMS, INC., CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:VOLKONSKY, VLADIMIR Y.;REEL/FRAME:009190/0032
Effective date: 19970826
|Nov 23, 1999||CC||Certificate of correction|
|Oct 25, 2002||FPAY||Fee payment|
Year of fee payment: 4
|Oct 13, 2006||FPAY||Fee payment|
Year of fee payment: 8
|Oct 6, 2010||FPAY||Fee payment|
Year of fee payment: 12
|Dec 11, 2015||AS||Assignment|
Owner name: ORACLE AMERICA, INC., CALIFORNIA
Free format text: MERGER AND CHANGE OF NAME;ASSIGNORS:ORACLE USA, INC.;SUN MICROSYSTEMS, INC.;ORACLE AMERICA, INC.;REEL/FRAME:037270/0159
Effective date: 20100212